Lab 09 - Unsupervised Classification

Earlier this semester we performed supervised classification when we created a Land Use/Land Cover thematic map from an aerial photo. In this lab, we created a thematic raster map using unsupervised classification-meaning we let the computer create the initial pixel classification groups, then we went in and refined them.

The lab started with an unsupervised classification of an .img file in ArcMap. Using the Iso Cluster tool to set the number of classes and iterations, and the Maximum Likelihood tool to set the rejection fraction option (to allow ArcMap to ignore outlying pixels that do not fit easily into the categories) I was able to quickly create an 8 category map. Although this was a simple method to create a thematic raster, there were errors in classification assignment-such as vegetation as water, etc.

To carry out a more exact unsupervised classification, I used the Unsupervised Classification function in ERDAS Imagine to create a raster with 50 classes, using 25 iterations and a convergence threshold of 0.950. This high number of iterations ensures the computer will continue refining until about 95% confidence (convergence threshold) of accurate pixel classification. I also set the skip factor to 2, for faster processing time, although this means the output will have pixels twice as large as the original.

Next was reclassification-walking through the output raster and changing the 50 classes into one of 5 classes-trees, shadows, urban, grass, or mixed. I experimented with different tools to compare the original image to the classified image to assist in accurate pixel assignment, preferring the swipe tool the most. The most challenging part of the reclassification was areas that had been grouped by the computer that were really two different groups, such as grass and buildings. So changing one affected changing the other. For some, I relied on whether the overwhelming majority of the pixels were in one class over the other, or I changed them to the mixed class if it was a fairly evenly divided split.

The last step was recoding and making my map layout. Recoding involved using the Thematic Recode function to select and recode the 50 groups to be only 5 groups. Easy enough, and from there I added a Class Name field and Area field to the raster's attribute table to calculate permeable vs. impermeable total areas before making my final map. While this was a straightforward lab, it is a bit challenging to deal with reorganizing the way the computer output the classes and dealing with the features being erroneously classified together. I am guessing that there are more steps that could break up those groupings, like cluster busting perhaps as discussed in our lecture.

Map of unsupervised classification in ERDAS Imagine with 50 classes which were reclassified and recoded into five groups.